acceptodds
Withdrawn from ICLR 2027

Regarding Multi-Instance Learning as a Denoising Process for Single-Cell Data

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Pooling single-cell embeddings into multi-sample bags can improve generalisation to unseen experimental settings. Canonically, bags are formed for convenience, since labelling a bag is cheaper than labelling every instance, and bags are generated the same way during training and at deployment, with instance-level labels free to vary within a bag (Dietterich et al.,1997). Dapello et al. (2025), by contrast, assume bags are constructed with perfect instance level accuracy such that each bag contains samples from the same class during both training and testing. In this work, we show that bag size and bag impurity act through distinct mechanisms. Bag size governs the variance of the mean embedding, with larger bags yielding lower variance, while impurity acts as a bias, shrinking the class-specific bag mean toward a shared centroid. This distinction motivates **Denoising DeepSets (DDS)**, which inserts a noise-conditioned denoiser between mean pooling and classification. By conditioning on the effective noise level implied by bag size, DDS can be trained across bag sizes and can be deployed at any evaluation bag size without retraining. On the datasets from the scGeneScope benchmark, DDS improves on mean-pooling and attention-based MIL baselines with the largest gains at small bag sizes where sampling noise is greatest. These results on both single-cell RNA sequencing and Cell Painting microscopy images suggest that treating bag size as a structured noise variable provides a principled route from bag-level training to robust single-cell prediction.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Multiple Instance Learning, Denoising, Perturbation Screens

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_qeNROpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper introduces the idea of multi-instance-learning for single cell-data, and combines it with a denoising-approach, which denoises bags instead of single instances.

Critical strengths

The paper introduces an interesting idea: That by combining cells into bags can improve performance.

Critical weaknesses

The paper is clearly identifying the problem, solution and results, and therefore should be rejected. In particular it is: (1) not describing the real-world application or concrete problem it aims to solve (2) not describing the method in enough detail, and (3) how the proposed solution compares to other baselines.

Minor comments:

  • The paper uses two citations in the abstract.
  • The paper is hard to follow. For example:
    • L. 033 The first sentence makes a statement without setting up any context.
    • L. 146 Introduces 'DeepSet model', 'k_train' and 'k_eval', 'macro recall', without introducing what dataset this is on, what a DeepSet model is, and how it is trained and evaluated. The authors could do a better job at carefully introducing the reader to concepts sequentially and building on each other.
    • L. 271. Mentions that the implementation supports 'removing a copy term', but this is not used. The relevance of the statement remains unclear.
Questions for authors
  • What is the motivation of introducing multi-instance-learning for single-cell data?

  • What is the downstream application the authors are targeting?

  • Which other methods are available that target this specific problem?

  • Could the authors describe the proposed algorithm, including potential training steps, with input and output variables?

Reviewer_r9DhOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper presents an approach to train a model for aggregating single cell data into bag profiles. The method distinguishes bag purity and bag sizes as parameters that can be modeled separately, and the study adopts the hypothesis that these can improve pooling performance. The method also introduces a denoising term to account for variance in bags and to improve purity and bag size stability. The proposed approach is validated in the problems of aggregating single-cell gene expression data and Cell Painting imaging data in the task of treatment identification.

Critical strengths

The modeling of bag purity and bag size is a unique angle to pooling single cell data. The introduction of the denoising component is an interesting addition.

Critical weaknesses
  • The paper evaluates the pooling of single cells through treatment classifiers. However, this task does not reflect a useful downstream analysis task for single cell data. Treatment is typically known experimentally as it is a choice made by the researchers. So having classifiers to detect treatment in single (or bags of) cells, is not useful on its own. It is a quality control check, but not a biological task. All the results and analysis are performed with classifiers for treatments, and misses the analysis of a real task such as mechanism of action clustering (for chemical perturbations).
  • The evaluation misses critical baselines, including simple averaging of the data on a fixed bag size regime, all the way to full treatments (meaningful if using a biological task rather than treatment classification). Other baselines such as a classifier conditioned on bag size, and a Set Transformer, among others, are missing. This leaves the current evaluation as superficial and limited to only a narrow set of methods.
  • The experimental results suggest that the denoiser is not as critical as suggested in the main text. When removing the denoiser and the shrinkage components, the performance drops only slightly and it is still superior to the mean pooling DeepSet (Table 11). This is not discussed or explained in the document.
  • The modeling of the statistics is not necessarily novel and it's mostly standard. While well motivated in theory, the method seems fragile in practice, especially considering that the improvements are marginal for a proxy task (treatment classification) rather than a core biological task.
Questions for authors
  • Why is treatment classification used as the task of focus? Why limiting the potential and impact of the methods to classifying treatments labels that are known to the experimentalist?
  • Can you add a biological task to evaluate the learned pooled representation?
  • Can you add more relevant baselines to compare performance?
  • Why does ablation indicate that the denoiser is not as impactful as stated in the main text?
Reviewer_K93yOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper investigates how the number and composition of single cell profiles aggregated into a bag impacts treatment classification. The authors differentiate between two sources of variation in this problem – sampling uncertainty due to smaller bag sizes, and bias introduced when bags contain cells from different treatment classes.

The paper has 2 main methodological contributions. 1) It uses finite population sampling theory to describe how the variance of a bag’s mean embedding depends on its size, and analyses how contamination shifts the expected representation. 2) It proposes Denoising DeepSets (DDS) which uses the estimated noise level associated with bag size to train a denoiser to reconstruct the full bag mean from a smaller sample before classification.

The authors validate the finite population variance relationship by comparing predicted and observed variance and investigate different bag size training schedules and levels of contamination. They evaluate DDS on scRNA-seq and cell painting data from scGeneScope benchmark - reporting improvements over mean pooling and attention based DeepSets, particularly for small bags and under held-out experimental settings. They also provide component ablations and an investigation of training stability.

Critical strengths
  • The statistical motivation is sensible and well justified. In particular, distinguishing between sampling uncertainty due to varying bag size and contamination bias is useful. The finite population analysis provides a principled way of estimating how sampling uncertainty changes with bag size, which is something the proposed method can potentially account for.

  • The experimental results are promising – DDS improves classification across both scRNA-seq and cell painting data, with the largest gains generally occurring for smaller bags, which is consistent with the authors’ motivation. It is also useful to see evaluations under experimental shifts, ablations of the architecture components and an evaluation of training stability.

Critical weaknesses
  • The contribution of the proposed learned denoiser is not completely clear. DDS combines existing approaches including DeepSets and EDM preconditioning with a learned denoising step. Although the full model performs best, the ablations in Table 11 suggest that the preconditioning potentially may contribute more to classification improvement than the learned denoiser itself (e.g. held-out macro recall decreases from 0.4954 to 0.4878 when the denoiser is removed, compared with 0.4700 when preconditioning is removed). The paper also does not directly report whether the denoiser improves reconstruction of the full bag mean, making it difficult to assess how much of the improvement is explained by the proposed denoising mechanism.

  • There appears to be discrepancies within some methodological and experimental details. Tables 3 and 6 report slightly different imaging results despite looking to describe the same setup. The appendix mentions that DDS and some baselines use different output class sets, which may affect the comparability of macro-averaged classification metrics. There also looks to be an inconsistency between Section 4.2 and Appendix D.3 regarding whether the signal scale is estimated during training or held fixed. These discrepancies make it harder to evaluate and reproduce the results.

Questions for authors
  1. In Table 11, the ablations suggest that EDM preconditioning may contribute more to the classification improvement than the learned denoiser itself. Held out macro recall falls from 0.4954 to 0.4878 when the learned denoiser is removed, compared with 0.4700 when preconditioning is removed. Could the authors provide further evidence of the benefit of the learned denoiser beyond preconditioning alone, such as whether the learned denoiser reduces reconstruction error relative to the full bag mean, and whether the improvements over the preconditioning-only model persist across different evaluation bag sizes?

  2. The appendix mentions that DDS and some baselines use different output class sets. Could the authors clarify whether the main comparisons in Tables 2 and 3 use identical treatment classes and test examples? Although macro recall and macro-F1 account for class imbalance, they do not make results from different classification tasks directly comparable.

  3. Tables 3 and 6 appear to report different imaging results for the same experimental setup. For example, held out macro recall for DDS at keval=1k_{\mathrm{eval}}=1 is 0.2334 in Table 3 but 0.2206 in Table 6, while under lognormal-k evaluation it is 0.3268 and 0.3386 respectively. Could the authors explain whether these results come from different configurations or clarify the discrepancy?

  4. Section 4.2 suggests that exponential moving estimates are maintained for both the within bag variability and σdata\sigma_{\mathrm{data}}, whereas Appendix D.3 states that σdata\sigma_{\mathrm{data}} is held fixed. Could the authors clarify which procedure was used in the reported experiments, and how the values of σdata=0.5\sigma_{\mathrm{data}}=0.5 for sequencing and 0.2310.231 for imaging were determined?

  5. The finite population correction in Equation 6 describes how sampling variance depends on bag size, but does not necessarily imply isotropic Gaussian noise as used to motivate the denoiser in Equation 2. Have the authors investigated how well this approximation describes the sampling error in the learned embeddings, particularly at small bag sizes such as keval=1k_{\mathrm{eval}}=1 where DDS reports its largest improvements?

Reviewer_qnPjOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the effect of bag size and bag impurity on multi-instance representations of single-cell data. The authors argue that reducing bag size primarily increases sampling variance, whereas mixing cells from different classes introduces bias into the aggregated representation. Based on this observation, they propose Denoising DeepSets (DDS), which applies a noise-conditioned denoiser to mean-pooled representations using a noise estimate derived from finite-population sampling theory. Experiments on the scGeneScope benchmark, covering both single-cell RNA sequencing and Cell Painting images, show improved performance across different bag sizes, particularly when only a small number of cells are available.

Critical strengths

(1) The problem is clearly motivated. The distinction between sampling variance caused by bag size and distributional bias caused by bag impurity is useful, and the statistical analysis provides a reasonable explanation for why models trained with large bags may perform poorly when evaluated with fewer cells.

(2) The improvements under small-bag settings are encouraging. DDS generally outperforms mean-pooling and attention-based DeepSets baselines when the number of available cells is limited. The evaluation of mismatched training and testing bag sizes also addresses a practical issue that is often overlooked in set-based single-cell modeling.

Critical weaknesses

(1) The methodological novelty is somewhat limited. The finite-population sampling variance is a standard statistical result, while noise-conditioned denoising and EDM preconditioning are adapted from existing techniques. The main contribution is therefore their application to pooled single-cell representations rather than a fundamentally new denoising framework. Furthermore, the ablation results suggest that EDM preconditioning accounts for much of the improvement, while the independent contribution of explicit noise conditioning is relatively modest.

(2) The experimental setting differs from conventional weakly supervised MIL. Bags are constructed using known treatment labels, meaning that cells within a bag are assumed to share the same treatment identity. This makes the task closer to supervised aggregation of cell profiles than classical MIL with unknown instance labels. In real single-cell applications, variation in cell types, treatment responses, and experimental conditions often occurs simultaneously. Modeling the uncertainty solely through bag size may be too restrictive for such settings.

(3) The generalization evidence is limited. Both modalities are evaluated using the same scGeneScope experimental framework, without validation on independent cell lines, perturbation collections, or laboratories. In addition, the Cell Painting experiments required several modality-specific adjustments, including control centering, fixed PCA representations, and modified training strategies. Although these preprocessing choices are also applied to the corresponding baselines, they make it difficult to draw a unified conclusion about the contribution of DDS across modalities.

Questions for authors

First, the authors should better isolate the contributions of random bag-size training, EDM preconditioning, and learned denoising. A particularly useful baseline would be a simple classifier that receives bag size or the estimated noise level as an additional input but does not include a denoising network.

Second, the method should be evaluated under more realistic changes in bag composition, experimental batches, and treatment conditions. The current held-out experiment setting primarily measures generalization across experimental rounds, rather than generalization to previously unseen biological perturbations.

These experiments would help determine whether DDS provides a broadly useful denoising mechanism or mainly improves robustness to a specific bag-sampling protocol.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.